Uploaded image for project: 'MariaDB Server'
  1. MariaDB Server
  2. MDEV-28994

Backup produces garbage when using memory-mapped log (PMEM)




      MDEV-14425 introduced a memory-mapped interface to the redo log when the log is stored in persistent memory, or in the /dev/shm file system. This improves performance and makes debugging more convenient, because the entire circular log file will be accessible at log_sys.buf in rr replay traces.

      However, this also exposed us to a (very reasonable) limitation of the rr debugger: it is assumed that the memory-mapped file will not be changed by processes that are not traced by the same rr record instance.

      Because it may be hard to remember the rule that mariadb-backup --backup must not be traced with rr record when the PMEM interface is being used, we’d better make backup refuse the use of the "fake PMEM" (mmap on redo log located in /dev/shm) when the process is apparently being run under rr record. Normal file system calls should be fine. We can use innodb_use_native_aio=0 as a proxy for detecting rr record, because both the old io_setup (libaio) nor the newer io_uring interface will return ENOSYS under rr record.

      Access to log files on real persistent memory (mount -o dax) will not be affected. It is assumed that rr record mariadb-backup --backup seldomly needs to be run such that the log is stored in persistent memory.


        Issue Links



              marko Marko Mäkelä
              marko Marko Mäkelä
              0 Vote for this issue
              3 Start watching this issue



                Git Integration

                  Error rendering 'com.xiplink.jira.git.jira_git_plugin:git-issue-webpanel'. Please contact your Jira administrators.